This document discusses collaborative filtering recommendation engines. It describes user-oriented and item-oriented collaborative filtering approaches. For user-oriented collaborative filtering, it finds similar users to a target user and generates recommendations based on the items those similar users purchased. For item-oriented collaborative filtering, it finds similar items to those a target user purchased and generates recommendations based on those similar items. The document also outlines challenges such as data sparsity, cold start problems, and scalability issues that recommendation engines may face.

1. RECOMMENDATION
ENGINE
DEMYSTIFIED
NEIGHBORHOOD METHODS
COLLABORATIVE FILTERING
Alex Lin
Senior Architect
Intelligent Mining

2. Outline
 Introduction
 User-oriented Collaborative Filtering
 Item-oriented Collaborative Filtering
 Challenges
 Best Practices

3. Recommendation Engine
What is a Recommendation Engine (RE)?
 RE takes “observation” data and uses machine
learning / statistical algorithms to predict
outcomes or levels of interest.
 “Recommender systems form a specific type of
information filtering (IF) technique that attempts
to present information items (movies, music,
books, news, images, web pages, etc.) that are
likely of interest to the user.” – Wikipedia

4. Recommendation Engine
Demographic
based
Clustering
Social Knowledge based
Graph based
based
Content
Collaborative based
Filtering Category
Mixture
Models based
 This
presentation will focus on
Neighborhood-based Collaborative Filtering
 User-oriented method
 Item-oriented method

5. Neighborhood-based
Collaborative Filtering
Data Normalization
Data
Input Data
(Observations) Representation
User-user similarity
vs.
Item-item similarity
Neighborhood
Formation
Make
Recommendation
Recommendation
Generation
Web Page, Email,
Direct Mail,
Marketing Campaign,
Delivering etc.
Mechanism

6. Outline
 Introduction
 User-oriented Collaborative Filtering
 Item-oriented Collaborative Filtering
 Challenges
 Best Practices

7. User-oriented CF
 Input Data Representation: Users / Items matrix.
users
1 2 3 4 5 6 7 8 9 10 … n
1 1 1 1 1
2 1 1 1
items
3 1 1 1 1 1
4 1 1 1 1
1 1
…
m
 Cell value “1” means user purchased the item.
 Data Normalization is not shown on this slide.

8. User-oriented CF
 Neighborhood Formation:
 Find the k most like-minded users in the system.
users
1 2 3 4 5 6 7 8 9 10 … n
1 1 1 1 1 1
2 1 1 1
items
3 1 1 1 1 1 1
4 1 1 1 1
1 1
…
m 1 1

9. User-oriented CF
 Neighborhood Formation:
 Find the k most like-minded users in the system.
users
1 2 3 4 5 6 7 8 9 10 … n
1 1 1 1 1 1
2 1 1 1
items
3 1 1 1 1 1 1
4 1 1 1 1
1 1
…
m 1 1
 Identify U9 and U2 are similar to U8

10. User-oriented CF
 Recommendation Generation:
users
1 2 3 4 5 6 7 8 9 10 … n
1 1 1 1 1 1
2 1 1 1
items
3 1 1 1 1 1 1
4 1 1 1 1
1 1
…
m 1 1
 Identify I1 and I9 are not yet purchased by U8

11. User-oriented CF
 Recommendation Generation:
users
1 2 3 4 5 6 7 8 9 10 … n
1 1 1 1 1 0.7 1
2 1 1 1
items
3 1 1 1 1 1 1
4 1 1 1 1
1 1
…
m 1 0.9 1
 Predict by taking weighted sum

12. User-oriented CF
Practical Implementation
 Compute and store all user-user similarities.
u•v
 Cosine similarity: sim(u,v) = cos(u,v) =
u ∗ v
2 2
 Find N items that will be most likely purchased by user u.
 Find k most similar users to u, save to Usim
 Get all items purchased by Usim, save to Icandidate
€
 Remove unavailable items in Icandidate
 Get all items purchased by u, save to Ipurchased
 Take Icandidate – Ipurchased = Irecmd
 Re-order items in Irecmd based on sum of user-user similarity
∑ v ∈k−similarUser(u)
userSim(u,v) * rvi
pred(u,i) =
∑ v ∈k−similarUser( u)
userSim(u,v)
€

13. Outline
 Introduction
 User-oriented Collaborative Filtering
 Item-oriented Collaborative Filtering
 Challenges
 Best Practices

14. Item-oriented CF
 Input Data Representation: Users / Items matrix.
users
1 2 3 4 5 6 7 8 9 10 … n
1 1 1 1 1
2 1 1 1
items
3 1 1 1 1 1
4 1 1 1 1
1 1
…
m
 Cell value “1” means user purchased the item.
 Data Normalization is not shown on this slide.

15. Item-oriented CF
 Neighborhood Formation:
 Find the k items that have the most similar user
vectors
users
1 2 3 4 5 6 7 8 9 10 … n
1 1 1 1 1
2 1 1 1 1
items
3 1 1 1 1
4 1 1 1 1 1
1 1
…
m 1 1

16. Item-oriented CF – cont.
 Recommendation Generation
1
W2-1
2 W2-3 3
Predict by taking weighted sum
4
TopN Recmd. for U8 : {1,9,3}
U8 4
W4-1 1
8
8
purchased W8-9 9
items
item-item
similarity

17. Item-oriented CF
Practical Implementation
 Compute and store all item-item similarities.
a•b
 Cosine similarity: sim(a,b) = cos(a,b) =
a ∗ b
2 2
 Find N items that will be most likely purchased by user u.
 Get all items purchased by u, save to Ipurchased
€
 For each item in Ipurchased, find k most similar items save them
to Icandidate
 Remove unavailable items in Icandidate
 Get all items purchased by u, save to Ipurchased
 Take Icandidate – Ipurchased = Irecmd
 Re-order items in Irecmd based on
∑ j ∈ purchasedItems(u)
itemSim(i, j) * ruj
pred(u,i) =
∑ j ∈ purchasedItems(u)
itemSim(i, j)

18. Outline
 Introduction
 User-oriented Collaborative Filtering
 Item-oriented Collaborative Filtering
 Challenges
 Best Practices

19. The Challenges you will face
 Data Sparsity Issue
 Cold Start Problem
 Curse of Dimensionality
 Scalability

20. Data Sparsity Issue
 Missing values in the Users / Items matrix.
users
Not
1 2 3 4 5 6 7 8 9 10 … n
Reliable Mostly
Unknown
1
2 ∑ userSim(u,v) * rvi
items
v ∈k−similarUser(u)
3 1
pred(u,i) = 1
∑ v ∈k−similarUser( u)
userSim(u,v)
4 1
1
…
m €
 NetflixPrize data set: 98.82% of cells are blank
 Typical e-commerce txn data set can be 10-100
time more sparse than Netflix Prize data set !!

21. Cold Start Problem
 It occurs when new item or new user is added to the
data matrix.
users
1 2 3 4 5 6 7 8 9 10 … n
1 1 1 1 1
2 1 1 1
items
3 1 1 1 1 1 1
4 1 1 1 1
1 1
…
m
 RE does not have enough knowledge about this new user or
this new item yet.
 Content-based REs can be incorporated to alleviate cold start
problem.

22. Curse of Dimensionality
 Adding more features (items or users) can
increase the noise, and hence the error.
 There aren’t enough observations to get good
estimates. users
items

23. Scalability
 User neighborhood formation: O(n2) for n users
 Item neighborhood formation: O(m2) for m items
 When m (# of items) << n (# of users), item-based
CF will be more efficient than user-based CF
 Ability to update neighborhood incrementally

24. Outline
 Introduction
 User-oriented Collaborative Filtering
 Item-oriented Collaborative Filtering
 Challenges
 Best Practices

25. Best Practices
 Understand the data thoroughly
 Define business objectives and conversion metrics
judiciously
 Understand context and user intent
 Apply adaptive reinforcement learning
 Optimize RE using cost-based methods
 Be aware of data-shift issue
 Optimize marketing messages delivered with
recommendation results

26. Understand the data thoroughly
 What data are available?
 E-commerce data set typically contains:
 Clickstream
 Shopping cart / Saved Items / Wish list / Shared Item
 Order / Return
 User profile
 User ratings
 How are these data points being collected?
 Is there pre-existing bias in the data? or leakage?
 Is the data related to what we want to predict?

27. Define business objectives and
conversion metrics judiciously
 Definingcorrect conversion metrics can be a
competitive advantage.
mitigate
increase
navigation
inefficiency
# of orders
up-sell increase
Web more
within average
property the context item price
revenue
cross-sell increase
within items
the context per order
Recommendation
Engine

28. Understand context and user
intent
 Context
should be considered when RE making
recommendation.
Month: December
Temperature: 45oF

29. Apply adaptive reinforcement
learning
C1
C2
C3
Incorporating clickstream
adaptive reinforcement
∑
j ∈ purchasedItems(u)
itemSim(i, j) * ruj
pred(u,i) = + W iC i
∑ j ∈ purchasedItems(u)
itemSim(i, j)
€

30. Optimize RE using cost-based
models
 Cost-based engine optimization
Metrics
Parameter

31. Be aware of the data-shift issue
 Data collection UI changes will influence data
significantly, creating artificial data shifting
Netflix prize data set
Y. Koren, "Collaborative Filtering with Temporal Dynamics," Proc. 15th ACM SIGKDD Int'l Conf. Knowledge Discovery and Data
Mining (KDD 09), ACM Press, 2009, pp. 447-455.12

32. Optimize marketing message
delivered with recommendation result
 It's How You Say It
Screenshots from Amazon.com

33. RECOMMENDATION
ENGINE
DEMYSTIFIED
Alex Lin
Intelligent Mining
Email: alin@intelligentmining.com
Twitter: DKALab